Non-Provisional Patent · Integrated Circuit

Ultra‑low‑power graph‑traversal on silicon.

A dedicated integrated circuit that walks a persistent typed event‑graph memory through hardware‑bounded arithmetic on decay‑adjusted link weights — within a milliwatt‑scale power budget built for wearables, vehicles, the home, and licensed SoC blocks.

mW‑scale power budget µs–ms traversal latency 8 principal blocks Host‑processor independent
FSM control FETCH LINK‑WEIGHT ALU DECAY CACHE RANK / SORT PWR REQ‑RESULT I/O MEM I/F

The architectural gap

Today's silicon can't walk a sparse graph at all-day battery life.

Personal AI assistants need a persistent memory they can recall in milliseconds. But the chips we have were built for other workloads — and they pay for the mismatch in power and latency.

// CPU

General-purpose processors

Evaluate every typed link in software — instruction fetch, memory access, and ALU cycles stacked per edge. The control overhead dwarfs the value of the weight produced.

// TPU / NPU

Tensor & neural accelerators

Tuned for dense, contiguous tensor math. Graph traversal is sparse, irregular, and data-dependent — so these engines sit idle while their datapaths starve.

// DATACENTER

HPC graph accelerators

Designed for graphs orders of magnitude larger, drawing tens to hundreds of watts. Wholly incompatible with a device you wear on your face all day.

// NONE OF THE ABOVE

No decay-aware hardware

None implement dedicated units for decay-adjusted link weights, decay-profile evaluation, hardware ranking, or traversal-aware caching. It all falls back to software.

Between watt-scale datacenter graph engines and dense-tensor wearable NPUs lies an architectural gap. The GTX-1 is purpose-built to fill it: a discrete, licensable block that does graph traversal — and nothing else — extremely cheaply.

Inside the boundary

Eight functional blocks. One traversal engine.

Every block is a fixed-function structure measured in pipeline cycles, coordinated by an on-chip state machine that runs while the host application processor sleeps. Tap a block in the floorplan to inspect it.

REQ‑RESULT MEM I/F 01Sparse‑Graph Fetch 02Link‑Weight ALU 03Decay Evaluation 04Candidate Ranking 05Traversal‑Control FSM 06Local Cache 08Power Management 07Request–Result Interface · IC boundary
The core innovation

Decay-adjusted link-weight arithmetic — in a single pipeline.

Five link attributes and an elapsed-time input are folded together through a decay function in a fixed-function datapath. What takes a CPU dozens of cycles, the GTX-1 resolves in one — or a handful.

confidence value link weight recurrence count temporal activation elapsed‑time input MAC multiply–accumulate f(decay) decay‑function apply NORM normalize decay‑ adjusted weight input reg MAC decay norm output reg

Five decay profiles, in fixed-function silicon

Each typed link names a decay-profile type; the decay-evaluation block routes it to the matching hardware function, parameterized by per-link coefficients.

Exponential

e−λt

Polynomial

1 / (1+t)k

Step-function

discrete tiers

Multi-segment

piecewise

Context-aware

state‑dependent
What lives in memory

The records the engine reads, ranks, and reinforces.

The persistent event-graph memory holds two record types. Highlighted fields feed directly into the decay-adjusted weight computation.

Typed Event Node

FIG. 11
node_ididentifier
node_type_idtype
attribute_fields[]payload
activation_magnituderanking
creation_timestamptime
last_access_timestamptime
access_countcounter
source_attributionprovenance
privacy_consent_meta[]policy
integrity_signaturecrypto

Typed Associative Link

FIG. 12
link_ididentifier
link_type_idtype
source_node / target_nodeedge
confidence_valueweight‑in
link_weightweight‑in
recurrence_countweight‑in
temporal_activation_valueweight‑in
decay_profile { type, coeffs }decay
last_traversal_timestamptime
privacy_consent_meta[] · sigpolicy
One block, many form factors

Licensable independently of any device.

The GTX-1 is framed as a discrete component — not an inseparable part of a product — so it can drop into wearables, cabins, homes, and third-party silicon alike.

Wearables

Smart-glasses, earbuds, smartwatches and other body-worn devices.

Vehicle-integrated

In-cabin AI, driver assistance, autonomous decision-support and fleet.

Home-integrated

Ambient assistance and occupant interaction in the connected home.

Industrial

Worker-context recall and operational decision-support on the floor.

Clinical

Clinician-facing patient recall under regulatory-compliance constraints.

Federated systems

Paired peer devices, each running an instance, sharing memory updates.

Chip-vendor SoC

Licensed to semiconductor vendors as a block of a mobile-processor SoC.

Privacy by boundary

Integrity-signed records, consent gating, raw sensor data never leaves.

Three ways to fabricate

Same protocol. Same structure. Any process.

The functional relationships hold across every manufacturing path — the engine speaks one request–result protocol regardless of how it's built.

01 / ASIC

Standalone die

A single semiconductor die on a dedicated process node, packaged with its own contacts to external systems.

02 / CHIPLET

Multi-chip module

Integrated as a chiplet through industry-standard interconnect, alongside other dies in one package.

03 / SoC BLOCK

Partition of a larger IC

A block within a general-purpose SoC sitting next to CPU cores, NPUs, modems, and crypto co-processors.

Many devices, one memory

Federated by design — without moving raw data.

FEDERATION HUB glasses earbud watch vehicle home

Attribute deltas, not sensor streams

Wireless transceivers — external to the IC — exchange federated update signals that carry only changes to link attributes, omitting any raw captured data.

A unified-view hub

The federation hub aggregates updates, distributes relevant changes per device-pairing and consent gating, and propagates right-to-erasure through tombstone and correction records.

Attestation-gated access

External AI agents reach memory elements only after authentication, cryptographic attestation of context, and consent-based authorization.

Hierarchical composition

Event-graph memories nest across individuals, groups, organizations, and consortia — each level with its own scope-governance policies.

Part of the memory subsystem

The accelerator that walks the graph — and the memory that holds it.

GTX‑1 does not store the event‑graph; it traverses one. Its dedicated memory interface is designed to pair with Priver Engram, a graph‑native compute‑in‑memory array that maintains the graph in place. Together they form the memory engine inside Priver Core: Engram stores and ages the links; GTX‑1 fetches, computes decay‑adjusted weights, and ranks candidates.

Priver Engram

The substrate

Stores the typed event‑graph; ages, grows, and reinforces links in place; returns decay‑adjusted weights and candidate nodes.

request

result
Priver GTX‑1

The accelerator

Walks the graph over the dedicated memory interface, computes link‑weight arithmetic in hardware, and selects what the inference layer sees next.

Abstract

A graph-traversal accelerator integrated circuit comprising sparse-graph fetch, link-weight arithmetic, decay-evaluation and candidate-node ranking circuitry, a traversal-control state machine, a local cache, and a request-result interface — all within an integrated-circuit boundary, coupled to a persistent typed event-graph memory through a dedicated memory interface. The link-weight arithmetic computes decay-adjusted weights as a hardware-bounded operation combining link weight, confidence, recurrence count, temporal activation, and elapsed time through a decay function. A power-management interface holds the part at milliwatt-scale through clock-gating, power-gating, dynamic voltage and frequency scaling, and duty-cycle management. Implementable as a standalone ASIC, a chiplet, or an SoC block.

22
FIGURES
20
CLAIMS
8
FUNCTIONAL BLOCKS
5
DECAY PROFILES